MKL_F16 type

zjin-lcf — Tue, 03 Sep 2024 19:19:59 GMT

Can you please explain if MKL_F16 is equivalent to the 16-bit floating-point type (i.e. half) ?

Running the example does not produce the expected output for MKL_F16.

https://github.com/zjin-lcf/HeCBench/tree/master/src/blas-gemm-omp

Thanks

Re: MKL_F16 type

Andrew_Barker — Thu, 19 Sep 2024 19:00:05 GMT

The MKL_F16 datatype is defined in the public header mkl_types.h to be the same as "unsigned short" (we do some conversions internally) so when you are setting the initial matrix and alpha, beta you are not getting the numbers you expect.

The simplest way to fix this in your linked example is probably to use _Float16 as your working type:

run_gemm_example<_Float16>(m, k, n, repeat);

and then cast the pointers to MKL_F16 in the actual gemm call:

 hgemm("N", "N", &n, &m, &k, (MKL_F16*) &alpha, (MKL_F16*) b, &n, (MKL_F16*) a, &k, (MKL_F16*) &beta, (MKL_F16*) c, &n);

topic Re: MKL_F16 type in Intel® oneAPI Math Kernel Library

MKL_F16 type

Re: MKL_F16 type